AI in RebateLedger: Every Number Is Computed, Not Generated

RebateLedger has fourteen live AI features, and one rule that governs all of them: the numbers are computed, not generated. Every threshold, score and rupee figure comes from a fixed formula running on your own data. The language model is only allowed to write the sentence around a figure it was handed — and if it writes a number the program did not produce, that explanation is rejected before you see it.

That boundary is the whole design, and it is worth understanding before the feature list, because it is what makes the feature list trustworthy.

The line between computing and explaining

Most software that claims to use AI is vague about which part of the answer the model produced. That vagueness is the problem: if a model wrote the number, the number carries the model's failure modes, and a finance team cannot defend it to an auditor.

RebateLedger splits the two jobs explicitly.

	What does it	What it produces
Computing	A deterministic program — SQL and statistics running on your own data	Every number, threshold, flag, score and rupee figure
Explaining	The language model	The plain-English sentence describing what those finished figures mean

Four things enforce that split, and each one is a fair question to ask any vendor:

• A guard-rail rejects invented numbers. The explanation is checked against the figures the program supplied. If it contains a number that was not in that set, it is thrown away and plain computed text is shown instead.
• Identifying data does not reach the model. What it receives is codes, identifiers and already-computed values — never partner names, GSTIN or phone numbers. Names are substituted back into the text after the model has returned.
• There are no AI forecasts. Forward-looking features run fixed formulas with documented floors. A forecast nobody can reproduce is not usable in a close.
• Every run is on the record. Each run stores its inputs, outputs, the explanation and a hash of the result, alongside the audit trail the rest of the product writes.

The practical consequence is simple. If you disagree with a figure, you can trace it to the data and the formula that produced it. The explanation is commentary on that figure, not the source of it.

The fourteen features

They cover cash recovery, risk scoring, forecasting and question-answering. All of them read data you already hold in RebateLedger — none of them needs a new integration or a separate data feed.

Feature	What it does
Margin Leakage Detector	Shows where margin is being lost — claims or schemes that have not been recovered
Recovery Aging Prioritizer	Ages overdue claims and settlements by how long they have been outstanding, like a receivables aging
Smart Opportunities	The rollup — recoverable and at-risk amounts across the workspace in one view
Anomaly & Duplicate Detector	Scans claims for outliers and duplicates that do not fit the normal pattern
Deduction Dispute Triage	Ranks un-decided inbound deductions and calls each one dispute, consider or accept
Slab-Optimization Recommender	Analyses how partner volumes land across an agreement's slab bands and flags structural problems, such as a band no partner falls into
GST Note Pre-Validator	Scans credit and debit notes for compliance issues before the return window
Agreement Summarizer	The live position on each sales agreement — booked volume, current slab and rate, money claimed, settled and pending
Purchase Agreement Summarizer	The same view for the rebates you receive from your principals and vendors
Counterparty Risk Scorecard	Credit and default-risk scoring across partners, computed from their activity with you
Period-Close Risk Scanner	Flags timeline risks and blockers that could delay your close, before they bite
Net Position Tracker	Your net settlement position and obligation forecast — what you are likely to owe and be owed
Festival-Season Demand Index	Reflects India's seasonal demand peaks, computed from your own history, so forecasting is realistic
Assistant	Answers questions about your own data in plain language

Two of these deserve a note, because their names could be read as promising more than they do.

Deduction Dispute Triage ranks inbound deductions and recommends a call; it does not dispute anything for you. The signal it uses is your own history with that counterparty — how past disputes with them were resolved — not a recomputation of your entitlement under the agreement. Thin history is scored against your own base rate and flagged as thin rather than presented as confident.

Slab-Optimization Recommender reports structural facts about an agreement's bands against actual partner volumes. It surfaces the problem; the commercial decision about whether to change a slab remains yours.

The step-by-step guides live in the Help Centre: finding recovery opportunities, detecting claim anomalies, reading risk and close insights, reading an agreement summary, pre-validating GST before settlement, the festival index and building what-if scenarios. The overview of how the layer fits together is in how Smart insights work.

What the assistant can and cannot do

The assistant answers questions about your own data — a claim's status, an accrual balance, a claims aging position, a settlement summary, your GST exposure, the health of a slab — and it explains decisions that have already been recorded, such as why a claim was decided the way it was, why a settlement is blocked, why a reconciliation shows a variance, or why a permission was denied.

Every one of those skills is read-only. The assistant cannot create, approve, edit, settle or delete anything. That is a property of how it is built rather than a policy someone could relax: the skills have no write path at all.

It is also narrower than it looks in one deliberate way. The model never sees your data. It sees your question and a catalogue of the things it is allowed to run, and it returns which one to run and with what parameters. The answer is then computed and written on the server. Piping your figures through a model to compose an answer would be a different architecture, and it is not this one.

The same discipline applies to documentation search: when it quotes the Help Centre, it quotes stored text directly rather than paraphrasing it, so the answer cannot drift from what the documentation actually says.

What happens when it cannot answer

It tells you, in fixed wording, and gives you a reference you can quote to support. The wording is not regenerated each time — it ships exactly as written, so an unanswerable question produces an honest reply rather than a fluent guess.

Questions it could not handle are recorded separately as coverage gaps, which is how the answerable set gets extended. A question refused because it was out of scope, and a question that failed, are not the same thing and are not counted together.

This matters more than it sounds. The failure mode people fear with an AI assistant over financial data is not that it says "I don't know" — it is that it says something plausible instead. Fixed fallback wording and a traceability check on every figure are the two mechanisms that prevent it.

On the roadmap, and not built yet

Four features are planned and not live: claim drafting from scanned documents, WhatsApp deduction capture, GST e-invoice IRN verification, and narration in additional languages.

They are named here because they are visible inside the product, where they appear marked "coming soon" rather than as working features. They are not part of the fourteen, and nothing on this page depends on them. If one of them is the reason you would buy, ask for a date before you do.

How to test any vendor's AI claims

The questions that separate real capability from packaging are short, and they work on any vendor, including this one.

1. Which part of the answer did the model produce — the number, or the sentence about the number? If the number, ask how it is validated.
2. What does it send to the model? Ask specifically whether customer names, GSTINs or transaction detail leave the system.
3. What happens when it does not know? Ask to see the failure, not the happy path. A demo that never shows a refusal is hiding one.
4. Can it change anything? An assistant that can approve or settle needs a much stronger control story than one that can only read.
5. Is every run reproducible? Ask whether you can retrieve what a given run computed, months later, and what it said.

The wider version of this test, applied to the category rather than to one product, is in can AI automatically match distributor claims to scheme terms, and the honest outlook for agent-driven settlement is in what is agentic claims processing. For the tax-document side specifically, see AI software for GST credit note automation.

Smart points you at what is worth attention; the recovery itself happens through the normal RebateLedger flows — raising the claim, reconciling it, settling it. That division is intentional, and it is why the AI layer can be switched off without stopping any of the work.